Experiments › Experiment Template

Notes · the whole page

Reader. A robotics researcher who has never seen your code. No run IDs, Hydra/YAML config names, W&B group names or CLI flags in the prose; say what each setup does ("the policy also sees fingertip contact forces"). Load writing:style before writing.
Title. Names what the experiment measures, never its result: "How much does fingertip tactile input help in-hand reorientation?" not "Tactile input doubles success". The homepage card uses the same title.
Language. English only, for now.
Write last. No standfirst and no Takeaways until right before publishing. The Takeaways sentence is also copied into <meta name="description">. Cards carry no text.
The draft is a dashboard. While runs go on, every running or planned figure is already on the page with its axes, legend and a status word in the middle (Running, Queued, Planned), and fills in when its runs finish. A figure whose runs are dropped is deleted. Publish only when no placeholder is left.
Order. The page opens with the cover animation alone: no heading, no caption, not in the contents. Then: Setup, Experiments, Takeaways, Related work, Acknowledgements, Experiment record, Citation. The side contents lists every section and, under Experiments, every question.
Cards on the homepage. Newest first; the grey template card stays last. A card is the cover picture, the date and the title, nothing else.

[Date of publishing, e.g. Nov 2, 2026]

[What the experiment measures, phrased as a question or a topic]

[Standfirst: empty while drafting. Right before publishing: two or three sentences with the answer, negatives included.]

By Qixuan Huang

Notes · cover

What it shows. A schematic of the setup, never a result. It carries the whole setup: the embodiment (arm, hand, DoF), the object, what the policy observes, what it outputs, and the one variable the page compares, so a reader can tell the experiment from the picture alone. A row of unrelated icons is not enough.
Format. Inline SVG, 1600×1000 (16:10), no text. If it moves: a 10–15 s seamless loop drawn on a canvas or SVG with a render(t) function, or an exported MP4 at 1920×1080 under 3 MB.
Palette. Navy ground #051E39; gold #B39051 for what moves or matters (the controlled links, fingers, the varied component); cream #F9F6E5 for inputs and observations (camera, object pose, touch); light gold #DEBD88 for goals and targets; slate #5f6b7a for fixed parts (table, base, walls). Flat fills, 2 px outlines, no gradients or shadows.
One style. Every other drawing on the page (setup panels, step-by-step explainers) uses this palette and line weight, with moving dots instead of arrows and as few labels as possible.
Card. The homepage card shows a still frame of the cover.

The two-link arm above is a stand-in.

Setup

[The baseline by its exact name, linked to the original paper and code at a fixed commit or release, e.g. "the Allegro-hand reorientation task from Isaac Lab v2.x, trained with PPO from rsl_rl". Then: what we change, and the metric. One or two paragraphs.]

baseline compute_reward() → ours (simplified)Code
  # simplified for reading; the Code link has the real code
  rot_dist = quat_angle(obj_rot, goal_rot)
- reward = -rot_dist
+ # [what this step does, in words]
+ reward = -rot_dist + [our term]

Notes · setup

Text. One or two paragraphs: the baseline by exact name with a link pinned to a commit or release; the embodiment (robot or hand model, DoF, sensors); simulator and version, physics step and control frequency; the task and what counts as success (e.g. "within 0.1 rad of the goal for 10 consecutive steps"); what we change; the metric. Define the compute axis once here (see Experiments). Sizes, step counts and seeds go in the figures, not here.
Diff. One diff of the core change and of any switch the figures compare, rewritten for reading: baseline lines red, ours green, both simplified alike, one comment per changed step, a header saying it is simplified, and a Code link to the real file.
Controls without code changes (domain-randomization ranges, number of parallel envs, action scale) go in the captions of the figures that use them.
Real robot. If any result is on hardware, say which, the robot and cameras, and the number of trials per setting.
End with a link to the exact public commit the runs used.

Notes · publishing the code

One branch per experiment on a public repo under github.com/k33ronng, named after the slug, carrying the history as it was run.
Before any public push, scrub cluster details from contents, file names and commit messages: PACE usernames, /storage/home/hcoda1/… and scratch paths, Slurm account and partition names, W&B entity names and API keys. Scan every file until nothing is left. Keep the unscrubbed original private.
A .gitignore block keeps credentials, wandb/, Slurm logs, checkpoints and raw videos out. Checkpoints and long videos go to a GitHub Release or Hugging Face, linked from the page.
Push only with the owner's approval. Code links go live after the push.

Experiments

Notes · experiments

By question. Each subsection is titled with the question its figures answer, in the order the questions came up, so the page shows how the work went. A tag before the title gives its status: AnsweredRunningQueuedPlanned
Answer first. An Answered question is followed at once by its answer: one or two sentences in words, at the level the figures support ("Yes, but only once the object is in grasp."), without supporting numbers. The tag alone is not enough.
Setup panels. Each subsection opens with a schematic, in the cover style, of the setups it introduces: one panel per setup, each with its name and one sentence. Tags mark the main setup Main and the one it is measured against Baseline. A setup drawn earlier is not drawn again.
Figure order. The first figure compares every setup on the raw metric with the baseline. Each later figure makes one point. Once the page settles on a setup, later variants follow that setup only.
Titles name what is plotted, never an axis or the comparison ("Reorientation success", not "Success vs. env steps against baseline").
Axes. Compute axis: environment steps summed over all parallel envs, or GPU-hours when wall-clock matters; say which. Success rate: the fraction of N evaluation episodes per seed (state N, and deterministic or stochastic actions), shown as the mean with a ±1 s.d. band over at least 3 seeds, 5 preferred.
Data, not typing. Charts are drawn from a generated data file (exported from W&B), never typed numbers, under the writing:plot rules. No running text lists numbers or restates a figure.
Captions state, in order: why this figure comes now, what is plotted, and what it shows in words. A placeholder's caption states the motivation and what it will test.
Videos. When a rollout shows something a curve can't (a slip, a regrasp, a recovery), add clips beside the figure: same camera and same seed across setups, muted, looping, ≤10 s, MP4 (H.264) with a poster frame, under 5 MB each, labelled by setup in the caption.

Running[First question, e.g. Does fingertip touch speed up learning in-hand reorientation?]

[Answer in one or two sentences once Answered.]

SETUP SCHEMATIC

[Setup name]Baseline
[One sentence on what it does.]

SETUP SCHEMATIC

[Setup name]Main
[One sentence on what it does.]

[Reorientation success]Running

Running · 3 of 5 seeds

0[Environment steps][max]

[Main][Baseline]

ROLLOUT · BASELINE

ROLLOUT · MAIN

[Why this figure comes now. What is plotted: success rate over N evaluation episodes per seed, mean ± 1 s.d. over 5 seeds, against environment steps. What it shows, in words.]

Planned[Question the results raised]

[What is plotted]Planned

Planned

[x-axis]

[Why this question came up, and what the figure will test.]

Takeaways

[Empty while drafting.]

Notes · takeawaysRight before publishing: one precise sentence with the result, negatives included. The same sentence goes into the page's meta description.

[Work, Authors Year]: [what it did, in one sentence]. [How this experiment relates to it: the same idea in a new setting, a test of its claim, or a different answer.]

Notes · related workThree to five of the closest works, not a survey: a link, what it did, and its relation to this experiment. Every reference is real and checked (open the paper; don't cite from memory).

Acknowledgements

[Thanks to people, names in alphabetical order. Delete this line if there are none.]

Thanks to Georgia Tech’s PACE team for the computing resources that made these experiments possible.

Notes · acknowledgementsThanks to people first, names in alphabetical order. Then the one PACE line, the same on every page; delete it if no PACE cluster was used.

Experiment record

Setup	[Varied variable]	Seeds	GPUs	Hours	Figures	Code	W&B
[Setup in words]	[value]	5	[1× H100]	[6.2]	1	[commit]	run
[Baseline in words]	[value]	5	[1× H100]	[5.8]	1	[commit]	run
Main runs: [X] GPU-hours ([GPU type]), [N] jobs
Other compute: [X] GPU-hours, [N] jobs

Notes · experiment record

No text above the table. One row per training run behind the page, generated by a script from the W&B export (finished runs) and Slurm sacct on PACE (queued and running jobs), never typed. Only main runs get rows.
Columns follow the topic: the setup in words, then each variable the page varies in a column of its own, then seeds, GPUs (count × type, e.g. H100, A100, L40S), hours, figures, code and W&B. No status column.
Two closing lines count from sacct over every GPU job of the project (GPU-hours = GPUs × elapsed): "Main runs" covers jobs that belong to a row, every resumed segment included; "Other compute" covers everything else: smoke tests, debug runs, superseded and abandoned runs. Together they are everything the page cost.
Links. Figures link to the figure on the page; Code to the public commit the run used; W&B links are absolute and open in a new tab. Every finished run links both. Never create a W&B page before the run exists.
Every finished run appears in at least one figure; a run in no figure gets a figure or leaves the table.
On phones no column is hidden: the setup column narrows and wraps, and the table fits at 375 px without sideways scrolling.

Citation

@misc{huang2026[slug],
  title        = {[Title]},
  author       = {Huang, Qixuan},
  year         = {2026},
  howpublished = {Blog post},
  url          = {https://qixuanhuang.com/exps/[slug].html}
}

Page format adapted from Wenhao Chai's Experiments.